Parallel Execution of Loops with Conditional Statements

This paper describes an approach to the evaluation of bounds of the execution time and number of processors needed to execute DO-like loops on MIMD systems. In the scope of this paper, we only consider single-nested loops. The evaluation of these bounds is done from information about dependences known at compile time. Due to the lack of information at compile time about the dynamic execution of the loop when conditional statements appear, an upper bound of the number of processors is obtained. In addition, a measure of the worst efficiency in terms of processor utilization is obtained by predicting the sequence of execution branches with the lowest parallelism. We also present some simulation results that show how the efficiency varies with the probability of the execution branches. Parallel code generation for shared-memory multiprocessors is outlined.